Procedia Computer Science Volume 29, 2014, Pages 1289–1298 ICCS 2014. 14th International Conference on Computational Science A Dataflow Programming Language and Its Compiler for Streaming Systems Haitao Wei, St´ephaneZuckerman, Xiaoming Li, and Guang R. Gao University of Delaware, Newark, DE, U.S.A [email protected], [email protected], [email protected], [email protected] Abstract The dataflow programming paradigm shows an important way to improve programming pro- ductivity for streaming systems. In this paper we propose COStream, a programming language based on synchronous data flow execution model for data-driven application. We also propose a compiler framework for COStream on general-purpose multi-core architectures. It features an inter-thread software pipelining scheduler to exploit the parallelism among the cores. We implemented the COStream compiler framework on x86 multi-core architecture and performed experiments to evaluate the system. Keywords: Dataflow Programming, COStream, Compiler, Streaming 1 Introduction As streaming systems based on multi-core processors have become ubiquitous, there is an urgent demand to design parallel programming models and compiler techniques to exploit parallelism on these systems. Parallel programming models like MPI and OpenMP provide a good way to perform parallel programming. But they still require the programmer to have the parallel model in mind and be careful to avoid data races, which adds to the burden of the programmer— especially for domain experts. Parallelizing compilers translate sequential programs to multi- thread ones automatically, but have only achieved limited success. Recently, Fresh Breeze [3, 4] and the Codelet Model [17], both rooted in dataflow, have proposed a promising execution and architecture model for dataflow programming to exploit pervasive models such as data, task and pipelining parallelism. SWARM [9] and DARTS [13] are two implementations of Codelet model, the former from and industrial, and the latter from an academic standpoint. They use a dynamic dataflow model (DARTS uses a hybrid static- dynamic dataflow model) to implement a fine-grained parallelism at runtime for the large scale parallel system. The data dependency is determined and scheduled at runtime. A significant amount of critical applications require a streaming model. They can be nat- urally supported by dataflow-inspired program execution models such as Fresh Breeze and Selection and peer-review under responsibility of the Scientific Programme Committee of ICCS 2014 1289 c The Authors. Published by Elsevier B.V. A Dataflow Programming Language and Its Compiler . Wei, Zuckerman, Li and Gao the Codelet Model. In the dataflow graph of streaming applications, each node is an au- tonomous computational unit. It has an independent instruction stream and address space and the data flow between nodes can be made through communication channels, implemented as FIFO queues. The dataflow model of computation exposes communications and an abundance of parallelism which offers the compiler many opportunities to lead to an efficient execution. The interesting thing is it uses the size of the data in each receiver/sender to control the ready signal to drive the computation. This can be used as a schema of rate control in the application. In this paper we propose COStream, a programming language as a implementation based on Flesh Breeze and Coldlet model. Compared with previous dataflow language like StreamIt [6], COStream adopts some grammar structure from IBM SPL [7] to improve the programbility and code reuse. It takes a multiple input/output actor instead of single input/output to support common computation. It uses explicit varible passing to make dataflow graph construction easy. It uses a special structure to support sub-graph construction which can be reused for larger graph. We also proposed a compiler framework for COStream. We implemented the compiler framework and evaluate it on x86 multi-core architecture. Section 2 presents the two program execution models we use as a fundation for COStream. Sections 3 and 4 respectively describe the COStream programming language and compilation framework. Section 5 presents our experimental results. Section 6 presents our vision to apply COStream to DDDAS. Finally, we conclude in Section 7. 2 Fresh Breeze and the Codelet Model 2.1 Codelet Model The Codelet Execution Model [17], is a hybrid model that incorporates the advantages of macro-dataflow [8, 5, 11] and the Von Neumann model. The Codelet Execution Model can be used to describe programs in massively parallel systems, which are expected to display a certain level of hierarchical or heterogeneity. The Codelet Execution Model extends traditional macro-dataflow models in the way shared system resources are managed. As in macro-dataflow, computation is done through units of small serial code known as codelets. The codelets are similar in their intent and behavior to the actors of macro-dataflow. However, where traditional macro-dataflow is only concerned with data availability, codelets can depend on other events, which refer to the availability of a (shared) resource, such as bandwidth availability, a specific processing element, a given power envelope, etc. Codelets are tagged with resource requirements and linked together by data dependencies to form a graph (analogous to a dataflow graph [2]). This graph is further partitioned into asynchronous procedures which are invoked in a control flow manner. This type of threading/synchronization model enables fine-grain execution. 2.2 The Fresh Breeze Memory Model The Fresh Breeze [3, 4] model also uses the concept of codelet as the core computational unit (i.e. event-driven non-preemptible tasks which are fired only when all dependencies are satisfied). In addition, its memory model uses trees of fixed-size chunks of memory to represent all data objects. Chunks are fixed size memory buffers (for now 128 bytes). Each chunk has a unique identifier, its handle, that serves as a globally valid means to locate the chunk within the storage system. Chunks are created and filled with data, but are frozen before being shared with concurrent tasks. This write-once policy eliminates data consistency issues and simplifies memory management by precluding the creation of cycles in the heap of chunks. Another 1290 A Dataflow Programming Language and Its Compiler . Wei, Zuckerman, Li and Gao benefit is that low-cost, reference-count based garbage collection may be used to reuse memory chunks for which no references exist in the system. Such a memory model provides a global addressing environment, a virtual one-level store, shared by all user jobs and all cores of a many-core multi-user computing system. It may be further extended to cover the entire online storage, replacing the separate means of accessing files and databases in conventional systems. 2.3 Discussion: Toward a Streaming Codelet Execution Model The stream programming model mentioned in this paper can be seen as a data-driven or stream implementation of Fresh Breeze/Codelet Model. In stream programming, the program is represented as a dataflow or stream graph. Nodes represent computation, edges represent communication implemented as FIFO queues. Following the Codelet model, computation is done by a codelet or an actor. Events in the stream programming model are triggered by the data items and FIFO buffer resources availability. The producer writes the data into the FIFO queue only when the computation is finished. The consumer reads the data from its FIFO queue at the beginning of the computation. Therefore, the write-once policy of Fresh Breeze Model is guaranteed by the single-writer / single-reader attributes of FIFO queues. 3 The COStream Programming Language 3.1 Operator and Stream Adopted from SPL [7], an operator in COStream is the basic computation unit. An operator can be seen as a procedure which consumes data from its input channel and produces data to its output channel. The data channel is called a stream. Different operators are connected by streams to construct a dataflow graph which represents the whole COStream program. An example of operator in Figure 1(a) is Aver, a component of the moving average application. Each operator contains three sections: init, work and window. Init is called at initialization time. In this case, Aver calculates the weights for the moving average. Work describes the most fine-grained execution step of the operator in the steady state schedule. Operators can be either stateful or stateless. A stateful operator depends on the local state of last execution and modifies the state at each execution. A stateless operator does not depend on the state of previous executions. The source in the example is a stateful operator, because the value of x in this execution depends on x’s value in the previous execution. 3.2 Window The only way an operator accesses the data from a stream is using windows. A window is bound to a stream. Within a work section, an operator can fetch data from the input stream window via a subscript in an array fashion, and put the result to the output stream window. As with SPL [7], there are two kinds of windows: sliding and tumbling windows. A sliding window has two parameters: the window size ws and the sliding size ss. A tumbling window only has one parameter: the window size. Each time a work section executes, the operator can read ws data items in the sliding window of input stream and slide out ss data items when the execution is done. The tumbling window is a special case of the sliding